\[\mbox{min}(\mbox{max}(x, 0), 1)\] instead of the sigmoid function (depolarisation constraints seem not work as they are likely to introduce some weird local minimum points).
SPR - 16 - C2 (Y: in red, R: in green, BG: in blue)
Here are some better results achieved by using a deeper model, and the finial activation function is replaced by
\[\mbox{min}(\mbox{max}(x, 0), 1)\] instead of the sigmoid function (depolarisation constraints seem not work as they are likely to introduce some weird local minimum points).
SPR - 16 - C2 (Y: in red, R: in green, BG: in blue)
SPR - 16 - D16 (Y: in red, R: in green, BG: in blue)
